The Slide Deck Is Now AI’s Problem
OpenAI's new ChatGPT add-in for PowerPoint asks users for the ability to read and modify documents, connect to Gmail, Outlook, and SharePoint, and send data over the internet. The permissions are listed in full on the Microsoft AppSource listing. What is not listed: what actually moves through those connections, where that data goes, or what OpenAI retains.
The app launched in beta on May 21, 2026, according to the AppSource listing, two days after Microsoft added GPT-5.5 to its own Copilot Chat feature, per Microsoft's own release notes. The competitive angle — OpenAI building a product inside Microsoft's own distribution channel, using Microsoft's own OAuth connections to reach the same data Microsoft uses — was covered by The Verge, Engadget, and Thurrott within hours of launch. The permissions disclosure received almost no coverage.
The AppSource entry lists the publisher as OpenAI, LLC and states: "When this app is used, it can read and make changes to your document and can send data over the Internet." The official ChatGPT app page confirms OAuth connections to Gmail, Outlook, and SharePoint, and describes the core features: creating new slides, updating existing decks, turning source material into presentation-ready content, and polishing slides directly in PowerPoint.
The product is available to free and paid ChatGPT users globally, including Business, Enterprise, Edu, and K-12 tiers. Per the app page, ChatGPT for Excel is generally available; PowerPoint remains in beta, with known limitations including complex formatting, template handling, native chart editing, animations, and custom fonts.
What the disclosure does not say is how those permissions are actually used. Standard OAuth integrations for Gmail and Outlook access typically require scoped permissions — limited to specific actions on specific data types — with explicit user consent for each. The AppSource language is notably broader: "read and make changes to your document" is not a scoped permission description, and "can send data over the Internet" covers a range of possible implementations, from necessary API calls to persistent telemetry.
The comparison that matters is Microsoft's own Copilot. Microsoft has published detailed documentation about what Copilot for Microsoft 365 accesses, how long it retains data, and what it does and does not send to OpenAI's servers for processing. That documentation exists because enterprise customers demanded it and because Microsoft's legal team understood the exposure. OpenAI's AppSource listing has none of that detail. A Microsoft spokesperson told Thurrott that the add-in went through Microsoft's standard AppSource review process. Microsoft did not address what data OpenAI retains or where it goes.
Anthropic shipped file-editing, including PowerPoint support, in Claude in September 2025, as Engadget reported. Claude's permissions model has faced its own scrutiny — the desktop app's file access prompted questions from security researchers about local file handling. Anthropic published a response detailing what the app accesses and under what conditions. OpenAI has not published equivalent documentation for the PowerPoint add-in as of this writing.
Enterprise IT teams evaluating the add-in will need answers to specific questions: Are prompts and document content sent to OpenAI's servers for processing, or does the model run locally? If prompts are logged, for how long, and under what access controls? Does OpenAI use data from these sessions to train future models? These are not exotic concerns — they are the baseline questions any organization with a reasonable security posture will ask before granting an add-in access to SharePoint and corporate email.
The product may be entirely benign. The OAuth connections may function exactly as a reasonable user would expect, and the "send data over the Internet" language may cover nothing more than the API calls required to render a slide. That is a plausible answer. It is also an answer that neither OpenAI nor Microsoft has provided publicly.
The beta has real constraints. Formatting limitations mean professional presenters will still need to touch slides manually. Enterprise deployment moves at IT approval speed. Microsoft's response to having its own ecosystem used as distribution for a competing AI product is not yet visible.
But the direction is clear. The friction of building a presentation — deciding what to say, structuring an argument, confronting the gap between what you meant and what the slides show — is the cognitive discipline the technology is now eliminating. That is a genuine shift, and it deserves a genuine answer about what the technology is actually doing with the material you give it to work with. Right now, that answer does not exist in a form that enterprise buyers or individual users can evaluate.